The news. On May 28, 2026, GrepSeek: Training Search Agents for Direct Corpus Interaction (arXiv:2605.29307, Salemi, Zeng, Diaz, Zamani et al.) trained LLM agents to interact with a text corpus through executable shell commands rather than a pre-built dense index. Training is two-stage: an answer-aware Tutor and answer-blind Planner generate verified search trajectories, then the policy is refined with GRPO. The paper reports the strongest token-level F1 and Exact Match across seven open-domain QA benchmarks, with a byte-exact parallel execution engine that speeds shell retrieval up to 7.6×. Read the paper →

Picture the rookie detective again. On day one she rummages through the case archive more or less at random — yanks open a drawer labelled "office," dumps a hundred folders on the desk, and can't say which line answers the question. That's an untrained agent firing a broad grep "office" at the corpus: lots of hits, almost all noise, wrong answer. Crucially, there is no card catalogue — no embeddings, no vector index to ask. The only way through the archive is to run a literal search and read what comes back.

GrepSeek's move is to coach the search itself. First a mentor who already knows each case's answer — the answer-aware Tutor — demonstrates the efficient sequence of drawer-pulls. Then the rookie, the answer-blind Planner, practises those moves without peeking at the answer, and the team keeps only the trajectories that actually cracked the case. That distilled set seeds a second stage of reinforcement learning: GRPO samples several command sequences per question, compares them against each other, and nudges the policy toward the ones that landed the right answer. Over training, the same agent stops grepping by habit and starts emitting a targeted pipeline — grep -i paris *.md | grep Q3 — that returns the handful of lines that matter.

Because the tool the agent calls is a literal shell, the win is two-sided. The retrieval is index-free, so there is no embedding pass or ANN store to build and keep fresh — the agent's tool is just a command line over the raw files. And the searches are learned end-to-end against the answer, so the policy adapts to this corpus and this question style rather than trusting a fixed similarity metric. This is the line worth drawing under the older agentic-retrieval results: where "Is Grep All You Need?" showed an untrained grep tool already competitive, GrepSeek shows what happens when you train the grep.

Where it sits among the options

Why the parallel engine matters for training

The byte-exact engine sounds like a systems footnote until you remember where RL spends its time. Each GRPO update needs many rollouts per question, and every rollout actually runs the agent's shell commands against the corpus. Say a single sequential grep sweep over the shards takes 760 ms (illustrative) and a training run does 100,000 rollouts: that is roughly 21 hours of pure retrieval before you count a single gradient step. The sharded-parallel engine runs those shards concurrently for a byte-exact identical result, collapsing 760 ms → ~100 ms — the same 100,000 rollouts now cost about 2.8 hours. The speedup is the real, reported 7.6×; the win compounds precisely because, in RL, you pay the retrieval cost on every rollout, not once.

Goes deeper in: AI Agents → Retrieval & RAG → RAG failure modes

Related explainers

• Is Grep All You Need? — grep vs vector retrieval for agentic search — the untrained-grep study GrepSeek builds on.
• VPO — vector reward vs GRPO — a closer look at the GRPO objective GrepSeek refines with.